zhaipro / unet

美发目标分割 + 染发展示
MIT License
30 stars 14 forks source link

秒记 #1

Open zhaipro opened 5 years ago

zhaipro commented 5 years ago

目标分割

zhaipro commented 5 years ago

https://blog.csdn.net/trent1985/article/details/80944942 https://gitee.com/onlysea/modihair/ https://keras.io/zh/layers/core/#permute http://vis-www.cs.umass.edu/lfw/ http://vis-www.cs.umass.edu/lfw/part_labels/ https://www.matrix.com/virtual-hair-color-try-on https://challenger.ai/dataset/lad2018 https://kazzkiq.github.io/svg-color-filter/ https://docs.rainmeter.net/tips/colormatrix-guide/ https://stackoverflow.com/questions/22999020/how-to-apply-svg-filter-to-group-using-svgwrite-with-python/23291406 https://yanweifu.github.io/papers/hairstyle_v_14_weidong.pdf https://github.com/lemondan/HumanParsing-Dataset https://github.com/TianzhongSong/Person-Segmentation-Keras https://github.com/bonlime/keras-deeplab-v3-plus http://www-scf.usc.edu/~liwenhu/SHM/database.html http://mmlab.ie.cuhk.edu.hk/projects/CelebA.html https://github.com/tensorflow/models/tree/master/research/deeplab 明度,亮度,HSL,HSV https://hanks.pub/2016/03/26/color-board/ https://open.mtlab.meitu.com/algorithm/imageProcessing/makeup

zhaipro commented 5 years ago
  1. 没有数据增强
    • 修改色相饱和度
    • 修改尺寸
    • 水平翻转
  2. 没有原始训练集
  3. 没有尝试更多模型
  4. 没有提供多种尺寸的图片用于训练,我想提升计算效率,达到在手机端能实时美发。
zhaipro commented 5 years ago

0.12, 0.93, 3.47M 700/700 [=] - 90s 129ms/step - loss: 0.0102 - dice_coeff: 0.9709 - psnr: 28.0166 - val_loss: 0.0110 - val_dice_coeff: 0.9679 - val_psnr: 27.5066

zhaipro commented 5 years ago

https://docs.python.org/3/library/zipfile.html https://stackoverflow.com/questions/18231135/load-compressed-data-npz-from-file-using-numpy-load

im = np.frombuffer(fp.read(), dtype='uint8')
im = cv2.imdecode(im, cv2.IMREAD_COLOR)
zhaipro commented 5 years ago

重要的灵感来源

Kaggle-Carvana-Image-Masking-Challenge 头发分割 曝光算法 HSL和HSV色彩空间

重要的失败经历

LUT 样条插值 开源的训练集可能还不如收费的预测集

不能再忘的基础知识

幂函数

zhaipro commented 5 years ago

距离美发还有很远的路要走,每天路边美女的发型告诉我。

zhaipro commented 5 years ago

在线展示

zhaipro commented 5 years ago

原本计划用染发效果来评价分割效果,但染发的挑战性超出了我的想象。

zhaipro commented 5 years ago

max(color) > 0 sum(mask) > 0

zhaipro commented 5 years ago

我的目标

zhaipro commented 5 years ago

mask = cv2.GaussianBlur(mask, (5, 5), 5)

zhaipro commented 5 years ago

我回来啦

zhaipro commented 5 years ago

从原图中提取明度,并做相应调整,结合上色的色相和饱和度。 现在尝试从原图中提取亮度,并做相应调整,保留原图的饱和度,结合上色的色相。 我们只是染发,修改的越少越好。

我们染发当然换了颜色,也提升了反光强度,让美发更明显,但饱和度是个什么鬼?

zhaipro commented 5 years ago
zhaipro commented 5 years ago

是因为我们头发亮度不充分的照片信息量太少了吗? 要不要单独训练一个模型,用于高清化?

zhaipro commented 4 years ago

美图给我的染发效果